1. MIMICKING THE HUMAN EAR Philipos Loizou (author) Oliver Johnson (me)

2. Anatomy and signal processing of the Ear Von Bekesy, max freq. response by location in Basilar membrane Implants are for people with hair cell loss causing profound deafness (90dB or more) and functioning cochlear nerve.

3. Auditory Encoding theories Place Theory Neural amplitude of firing differs by location for different frequencies. Brain decodes combination of amplitudes’ location of firing. Motivates multichannel implants. Volley Theory, frequency determined by neuron firing rate. Nerve fibers fire at rates proportional to period of input signal, up to 5kHz. Low frequencies nerve fibers fire at each cycle, high frequency is encoded by combinations of groups. Is this not spatially dependent?

4. Speech Encoding Source-filter model of speech production Linear system of F0, F1, F2, F3

5. Cochlear Implant Design Transmission Link Electrode Design Stimulation Type Signal Processing

6. Electrode Design Typically implanted near the scala tympani to be near the auditory neurons that lie along the length of the cochlea (preserves ‘place’ mechanism Bipolar vs. Monopole Insertion depth Up-shifting in frequency when F0 is represented at 22mm depth

7. Single Channel vs Multi-Channel Author surprised at success rate for single channel implants. Single Channel Limited spectral information Signal contains information on F0 and F1, some patients could discriminate F2 Used temporal encoding only up to 1kHz (neural refractory) Can extract freq information from periodicity of input stimulus up to 300- 500Hz Multi How many electrodes to use What information should be transmitted, what does the ear encode

8. Stimulation Types – Waveform Strategies Compressed Analog and Continuous Interleaved Sampling Approaches Analog Advantages: Contiguous frequency bands, simpler electronics Disadvantages: Cross channel interference Pulsatile Advantages: Increase pulse rates, more control over stimulation ordering, can envelope signal to person’s dyn. range. Implant listeners range as low as ~5dB Disadvantages: Required digital impl.

9. CA and pulsatile comparison

10. CA vs CIS

12. Electrode placement and mapping

13. Speech Decoding – Feature Extraction MPEAK saying “sa”

14. Speech Decoding SMSP displaying pronunciation of the word “choice”

15. Speech Decoding SPEAK Pronunciation of several phonemes.

16. Hybrid Systems Spectral Maxima Sound Processor (SMSP) Combines CIS components with more electrodes doing feature extraction for spatial mapping Electro-Acoustic Systems Low freq hearing aid amp. Cochlear implant for high frequencies

17. Improvement in speech and hearing

18. Factors Affecting performance Duration of Deafness Age of onset People with post-lingual deafness perform better than those with pre-lingual deafness Age at implantation Possibly minimum age of 2, better young than older. Duration of implant use Individual Factors Number of surviving spiral ganglion cells Electrode placement and electrode depth Electrical dynamic range Signal processing strategy

19. Pre- vs. Post-lingually deafened performance comparison

20. Implantation timeline

21. Conclusions (look at more recent papers) Proven that profoundly deaf people can hear and speak with the implant. Longitudinal studies have shown this to be true for a majority of pre-lingually deaf as well Auditory performance is most positively affected by two factors Acquiring speech and language before hearing loss Shorter duration of deafness Signal processing advances in both speed of processing and speech decoding

22. Future work Continued comparison of feature extraction vs. CIS methods. Discover more Inter-patient performance variability factors Develop tests to predict a person’s performance pre- surgically, such as the P100 look at visual cortical re- mapping. Electrode array design for high degree of specificity to allow finer channel selectivity

23. Questions raised… Comparison of what processing strategies work better for pre- vs. post-lingual deafness onset Longitudinal studies for implantation success (one study with 30 people only had 9 electrode replacements in 8 people) over a 10 year study A lot of manual refinement by audiologists, did not see much in terms of optimizing algorithms/adaptive learning As a model for other BMIs for how long it takes a person to truly integrate with a machine interface Initiate Real discussions on ‘elective’ BMI procedures, gives us a roadmap for future applications. Spoken language differences